--- title: "Which PixAI Model Should You Use? A Guide for Every Task" description: "Tsubaki.3 is the default answer for most PixAI models work, but not all of it. A task-by-task guide to picking a model, including where Tsubaki.3 loses." canonical: "https://blog.pixai.art/en/pixai-models-guide/" language: "en" published: "2026-09-13T03:00:00+00:00" last-updated: "2026-09-02T08:07:31+00:00" author: "PixAI Official" categories: "Models & LoRAs" image: "https://images-ng.pixai.art/images/orig/d46a5743-bf8c-4b14-b22f-ca21a081827e" site: "PixAI Blog" publisher: "Mewtant Inc." --- # Which PixAI Model Should You Use? A Guide for Every Task MODEL GUIDE PIXAI BLOG PIXAI.ART 2026 — WHICH MODEL TO USE — Which PixAI Model Should You Use? Dozens of models, one decision. This guide goes task by task — and the two questions that settle most of it are simpler than the model list suggests. [▸ Open the PixAI Generator](https://pixai.art/generator/image?utm_source=blog) > Model usage numbers and interface screenshots reflect September 2026. Same prompt. Same composition instructions. The only variable is the model. ![PixAI Tsubaki.3 result: two girls in a cafe](https://images-ng.pixai.art/images/orig/37c0c314-b433-4ffd-a828-e1ffa9d3f901) Tsubaki.3 ![PixAI Haruka v2 result from the same prompt](https://images-ng.pixai.art/images/orig/c410ef41-46df-4fcc-b5d5-fea38600eb8c) Haruka v2 On the left, the cream cardigan is on the braided brunette and the grey blouse is on the purple ponytail, exactly as written. On the right the two outfits have swapped, both girls have their eyes closed, and both are holding a book. **Neither result is better than the other. They are good at different jobs.** PixAI has dozens of official models, and the community LoRA library on top of that is effectively bottomless. The short version: **start with Tsubaki.3, and switch systems only where it falls short.** This guide is about where exactly it falls short, sorted by what you are trying to draw. Three parts. The first four chapters settle the big decisions, the middle seven go task by task, the last two cover the full lineup and cost. If you only want the answer, the table in Chapter 01 is enough. If you want to follow along, keep the [generator](https://pixai.art/generator/image?utm_source=blog) open in another tab. — PART ONE — Before you start | Two things to settle first — CHAPTER 01 — ## The 30-second answer The answers first. The reasoning is in the chapters below. | What you want to make | Start with | Also consider | | --- | --- | --- | | One character, simple composition | **Hoshino v2** | Haruka v2, or Tsubaki.3 to describe the scene in a sentence | | A long, multi-clause prompt | **Tsubaki.3** | Tsubaki.2 | | Two or more characters in one frame | **Tsubaki.3** | Haruka v2 + multi-LoRA | | Manga panels, text inside the image | **Tsubaki.3** | Tsubaki.2 | | The same character in new scenes | **Tsubaki.3** with a character reference | Haruka v2 + character LoRA | | Editing an existing image from a reference | **Reference Pro** | Edit Pro for multi-part instructions | | A specific art style | **Haruka v2** + community LoRA | Hoshino v2 / Otome v2 | | Soft, low saturation, romantic | **Otome v2** | Crystalize / Eternal | | Sharper features, more mature | **Hoshino v2** | Nagi | | Korean webtoon look | **Serin** | Haruka v2 + style LoRA | | Male characters as the subject | **Otome v2** | Otome v1 | | Lowest credit cost | **Tsubaki Flash** | Haruka v2 | | Photoreal rather than anime | **PixAI Real V4.0** | Tsubaki.3 | Two columns, because most jobs have a fast answer and a second option worth knowing. If your case is not in the table, the chapters below go through them one at a time. — CHAPTER 02 — ## Models and LoRAs are not the same thing Worth settling before anything else, because everything after this depends on it. > **A model is the foundation. It decides whether the image can be made at all. > A LoRA sits on top. It decides who it looks like, or what style it is in.** You pick one model per generation, and you can stack several LoRAs. LoRAs are tied to the architecture they were trained for, so the LoRA you want can end up deciding the model — which is why Chapter 04 is worth reading before you commit to anything. For more: [Model vs LoRA Foundations](https://blog.pixai.art/en/model-vs-lora-pixai-foundations/) and [What Is LoRA in AI Art](https://blog.pixai.art/en/what-is-lora-beginners-guide/). Stacking rules are in the [LoRA Stacking Guide](https://blog.pixai.art/en/lora-stacking-pixai-guide/). — CHAPTER 03 — ## What Tsubaki.3 is good at [Tsubaki.3](https://pixai.art/model/2024383378759147749?utm_source=blog) is the newest generation of the Tsubaki family. Three things it does that the tag-driven models do not: 1. **Plain sentences go through.** No tag syntax to learn. “Two girls by a window, the one on the left resting her head on the other’s shoulder” is a valid prompt, in English or not. 2. **Multiple subjects stay separate.** Write “A has blue hair, B has a brown braid” and the traits attach to the right person. This is where tag-only systems break down hardest. 3. **Text inside the image is readable.** Short English lines in a speech bubble come out correctly spelled. For manga panels or signage, this family is effectively the only option (more below). ![PixAI Tsubaki.3 result: a girl at a rooftop railing at golden hour](https://images-ng.pixai.art/images/orig/90831ea5-b9b3-47cc-bac8-cc4dac1870be) Object placement, light direction and background were all specified in plain sentences Coming from [Tsubaki.2](https://blog.pixai.art/en/tsubaki-2-is-here-everything-you-need-to-know/)? If you write long prompts, draw multiple characters, or need text in the image, moving over is a clear win. If your prompts are short and you already have a look you like, there is no urgency. — CHAPTER 04 — ## DiT and SDXL: what each family is good at PixAI models split into two architecture families that read instructions in fundamentally different ways. This is the single most useful thing to know before picking anything, because **LoRAs only work with the family they were trained for.** ### DiT (Tsubaki.3, Tsubaki.2, Tsubaki, Serin) Reads sentences as sentences. **Relationships inside a long clause survive** — “to the left of”, “while doing” — and multiple characters and in-image text belong to this family too. A real prompt opens with `artist:` or `character:` tags and then describes the scene in ordinary prose. ### SDXL (Haruka v2, Hoshino v2, Otome v2, Hinata v2, Nagi, Crystalize, Eternal) Reads a tag list. **The finest style granularity, and an order of magnitude more community LoRAs** — if you want a specific brushwork or a specific artist’s feel, the answer is almost always here. The trade: long prompts get truncated, tags are not bound to each other, and traits bleed between characters. **The practical consequence:** almost all community LoRAs on PixAI were built for SDXL. If your starting point is “I want to use this LoRA”, check which family it belongs to first — that decision comes before the model. The two images at the top came from these two prompts. ``` [DiT — Tsubaki.3] character: original, A quiet cafe by the window on an overcast afternoon. The girl on the left, with short brown hair in a braid, leans sideways and rests her head on the other girl's shoulder. The girl on the right, with long purple hair in a ponytail, keeps her eyes down on the book in her lap. A half-finished latte sits between them. [SDXL — Haruka v2] 2girls, cafe, window, overcast, brown hair, braid, purple hair, ponytail, leaning on shoulder, reading book, latte, soft lighting, masterpiece, best quality ``` The tag list has no way to say which girl is the brunette — that is the boundary line. Run it the other way and the prose prompt cannot reach the style granularity SDXL gives you. Writing details for each: [DiT Model Prompt Writing Guide](https://blog.pixai.art/en/dit-model-prompt-writing-guide/) and [prompt guide for SDXL models](https://blog.pixai.art/en/general-guide-of-writing-prompts-for-sdxl-models/). — PART TWO — Pick by the job | Seven situations — CHAPTER 05 — ## One character, simple composition The most common case by far. Standing pose, bust shot, simple background. Either family handles this well, and the SDXL models are the practical default here. [Hoshino v2](https://pixai.art/model/1954632827019711809?utm_source=blog) or [Haruka v2](https://pixai.art/model/1861558737426484240?utm_source=blog) are lighter, with less variance between generations — which matters when you are rolling for a winner. Reach for [Tsubaki.3](https://pixai.art/model/2024383378759147749?utm_source=blog) when you would rather describe the scene in a sentence than assemble a tag list. New to PixAI? Start with [How to Use PixAI](https://blog.pixai.art/en/how-to-use-pixai-guide/). Differences between the SDXL models are covered in [SDXL Anime Models on PixAI](https://blog.pixai.art/en/sdxl-anime-models-pixai-guide/). — CHAPTER 06 — ## Long, complex prompts This is where Tsubaki.3 wins most clearly. Direction, action, props, where the light falls — you can hand over the whole thing at once. ``` character: original, A medium shot from a slightly low angle. A girl in a navy sailor uniform stands at the top of an outdoor staircase, holding a folded umbrella in her right hand while her left hand grips the railing. Late afternoon light comes from the left, so the left side of her face is brightly lit while the right side drops into shadow. Cherry petals drift past her from behind. She looks down toward the camera with a faint, tired smile. ``` ![PixAI Tsubaki.3 result from a long prompt: a girl on an outdoor staircase](https://images-ng.pixai.art/images/orig/9e819771-0334-421a-aa55-0846730c6b63) Umbrella in the right hand, left hand on the railing, light split across the face — all as written Convert that to a tag list and there is no way to say which hand holds the `umbrella` and which grips the `railing`, or which side the light comes from. You get umbrellas in both hands, or the shadow on the wrong side. **Tags are not bound to each other**, and no amount of prompt rewriting fixes that. How to write long prompts: [The PixAI Prompt Formula](https://blog.pixai.art/en/how-to-write-pixai-prompts-formula/). Steps, CFG and the rest: [PixAI Advanced Settings](https://blog.pixai.art/en/pixai-advanced-settings/). — CHAPTER 07 — ## Choosing an art style When the starting point is a look rather than a scene, you are choosing a style, not a model — and the style inventory lives in SDXL. | The look you want | Model | | --- | --- | | Soft, low saturation, romantic | [Otome v2](https://pixai.art/model/1856956427276171616?utm_source=blog) | | Sharper features, more mature | [Hoshino v2](https://pixai.art/model/1954632827019711809?utm_source=blog) | | Balanced, meant for stacking LoRAs | [Haruka v2](https://pixai.art/model/1861558737426484240?utm_source=blog) | | Korean webtoon | [Serin](https://pixai.art/model/1911477338694141481?utm_source=blog) | | Clean, cooler palette | [Crystalize](https://pixai.art/model/1964824607980730248?utm_source=blog) / [Eternal](https://pixai.art/model/1965302325046914604?utm_source=blog) / [Nagi](https://pixai.art/model/1961253300562760878?utm_source=blog) | Same prompt, same seed, only the model changed: ![PixAI Otome v2 portrait of a girl in a school uniform](https://images-ng.pixai.art/images/orig/3cdf3edb-4bdc-4507-acb6-fa5f24b3aef3) Otome v2 ![PixAI Hoshino v2 portrait of a girl in a school uniform](https://images-ng.pixai.art/images/orig/cae05501-b2cc-4276-9e7d-1a0048033578) Hoshino v2 ![PixAI Haruka v2 portrait of a girl in a school uniform](https://images-ng.pixai.art/images/orig/0d789721-1f53-40ea-b4c3-d3d79879a207) Haruka v2 ![PixAI Serin portrait of a girl in a school uniform](https://images-ng.pixai.art/images/orig/d9b7d330-e7aa-481a-8dfa-66580afcae37) Serin To push further, stack an official style LoRA. [More Natural Images](https://pixai.art/model/1991224148554349023?utm_source=blog) pulls linework and shading toward natural, [Punchy Anime Plus](https://pixai.art/model/1991223785685907182?utm_source=blog) lifts saturation and contrast, and [Haruka v2 style](https://pixai.art/model/1991205934961235887?utm_source=blog) is an easy starting point. Anything more distinctive is in the [model market](https://pixai.art/market?utm_source=blog). On the Tsubaki side, style is locked with [Customize Style and Style Code](https://blog.pixai.art/en/customize-style-feature-guide/) rather than LoRAs. You can aim in a direction, but you cannot get SDXL-level precision. That is the honest gap today. ### If your subject is a male character Most anime models lean toward female characters by default. For male characters, start with [Otome v1](https://pixai.art/model/1854274014747848600?utm_source=blog) or [Otome v2](https://pixai.art/model/1856956427276171616?utm_source=blog), which are designed specifically for male-focused generation. Tagging matters more than the model here: [Mastering Male Characters](https://blog.pixai.art/en/pixai-sdxl-prompting-guide-foundations-mastering-male-characters/) and the [male character tag cheat sheet](https://blog.pixai.art/en/pixai-sdxl-prompting-guide-foundations-male-character-tag-cheat-sheet/). Models that suit josei-oriented work are listed [here](https://blog.pixai.art/en/pixai-sdxl-prompting-guide-foundations-recommended-otome-friendly-models/). — CHAPTER 08 — ## The same character, or several in one frame The short answers: - **You have a reference image** → [Reference Pro](https://pixai.art/model/1948514378379046913?utm_source=blog). Fastest route. - **You need dozens of images over months** → train a character LoRA. Higher setup cost, most stable result. - **Two or more characters in one frame** → Tsubaki.3. List the `character:` tags and describe each one separately in prose. Each has a dedicated guide: [PixAI Character Consistency](https://blog.pixai.art/en/pixai-character-consistency-3-beginner-methods/), [the Reference Pro guide](https://blog.pixai.art/en/pixai-reference-pro-guide-multi-image-editing-with-natural-language/), and [Multi-Character LoRA Generation](https://blog.pixai.art/en/multi-character-lora-generation-guide/). — CHAPTER 09 — ## Manga panels and text inside the image Speech bubbles, signage, title lettering. **Only the Tsubaki family produces readable text.** SDXL gives you scrambled glyphs. The trick is to describe the bubble as an object you want drawn. Shape, outline, tail direction, and the lettering inside. ``` one single white rounded speech bubble with a black outline and a short tail pointing toward the girl on the left. Inside it, in clean bold black comic lettering, perfectly spelled, is the text: "You could've helped." [add to the negative prompt] two speech bubbles, empty speech bubble, garbled text ``` ![PixAI Tsubaki.3 result: a comic panel with a speech bubble](https://images-ng.pixai.art/images/orig/e59a8f06-dda9-422d-996e-5d0b8c73912a) Short lines come out correctly spelled, apostrophe included The honest limit: **short English lines are reliable, long sentences and non-Latin scripts are not.** For anything longer, have the model draw an empty bubble and add the lettering afterwards in an image editor. — CHAPTER 10 — ## Editing an image you already have Not a generation-model question. The confusing part is that there are four entry points. | Tool | How you give instructions | Best for | | --- | --- | --- | | [PixAI Edit](https://pixai.art/model/1882220606138613945?utm_source=blog) | Instruction + one image | Standard edits. Start here | | [PixAI Edit Lite](https://pixai.art/model/1952119659421483010?utm_source=blog) | Same, lighter | Many small edits in a row. Found under the MARKET tab in the model picker | | [PixAI Edit Pro](https://pixai.art/model/2006468692900798466?utm_source=blog) | Instruction + images | Complex instructions, several changes at once | | [Flow Edit](https://pixai.art/generator/chat-editing?utm_source=blog) | Conversational, in sequence | Stacking “a bit brighter” over and over | ![PixAI Edit interface with one reference image and a one-line instruction](https://images-ng.pixai.art/images/orig/e4b46b2e-4643-487a-8470-3429d9e72881) PixAI Edit: one reference image, one line of instruction ![PixAI model picker MARKET tab showing PixAI Edit Lite](https://images-ng.pixai.art/images/orig/35de7b43-38ab-4812-af70-82376f173641) Edit Lite is not in PRESET. Switch to the MARKET tab to find it ![PixAI Edit Pro interface with two reference images and a multi-part instruction](https://images-ng.pixai.art/images/orig/f723b4b5-8a5d-4157-86b4-4ab73c77340b) Edit Pro: background, outfit and prop changed in one sentence ![PixAI Flow Edit conversational editing workspace](https://images-ng.pixai.art/images/orig/11b2c55f-c1ec-4e39-a932-1438491dcda0) Flow Edit: “make it a bit brighter”, then “blur only the background” Fill, partial repaint and upscaling are a different job — you are selecting a region rather than giving an instruction. That is covered in [the inpainting and outpainting guide](https://blog.pixai.art/en/ai-image-inpainting-outpainting-guide/). If this makes you want to build your own model, [Train LoRA on PixAI](https://blog.pixai.art/en/train-lora-on-pixai/) and the [DiT LoRA Training Guide](https://blog.pixai.art/en/dit-lora-training-guide/) are the entry points. — CHAPTER 11 — ## Choosing on cost and speed The second axis, after style. A day spent on one image and a day spent rolling for a winner call for different answers. - **One image, carefully** → whichever model best fits the job from the chapters above. The more detailed the instruction, the more a DiT model pays back - **Volume** → [Tsubaki Flash](https://pixai.art/model/1934789863877919615?utm_source=blog), or SDXL - **Running low on credits** → work out the composition on SDXL first, then rebuild only the keeper on Tsubaki.3 Spending the expensive step only on a composition you have already committed to is what keeps the credit burn slow. Ways to top up: [the free credits guide](https://blog.pixai.art/en/pixai-free-credits-guide/). --- — CHAPTER — ## FAQ ### Which PixAI model should beginners start with? It depends on your goal. Try **Tsubaki.3** for natural-language prompting and complex scenes, **Haruka v2** for flexible SDXL styling and LoRA use, or **Serin** for a Korean webtoon-inspired look. ### When should I use an SDXL model instead of Tsubaki.3? Choose an SDXL model when the style or LoRA matters more than complex instruction following. PixAI’s SDXL ecosystem offers a wider range of community LoRAs and more granular style choices, while Tsubaki.3 is better suited to long prompts, spatial relationships, multiple subjects, and text inside images. ### Can I use LoRAs with every PixAI model? No. LoRAs are tied to compatible model architectures. An SDXL LoRA needs a compatible SDXL base model and cannot be loaded directly with DiT models. If a particular LoRA is essential to your workflow, check its compatibility before choosing your base model. ### Which PixAI option is best for keeping the same character consistent? It depends on what you are making. For generating the same character in new scenes, Tsubaki.3 with a character reference is a good starting point. If you are editing an existing image, Reference Pro is designed for reference-based editing. For a character you plan to reuse across many images over time, training a character LoRA may be the more scalable option. ### Which PixAI model is best for multiple characters, manga panels, or text inside images? Tsubaki.3 is the strongest starting point for these tasks. It handles relationships between multiple subjects more reliably than tag-based SDXL workflows and is better suited to manga-style layouts and short text inside the generated image. For longer dialogue or more precise typography, it is still better to add the final text during editing. --- Model choice looks like a lot to memorise, but two questions settle most of it. **Do you need a specific community LoRA?** That decides the architecture family before anything else. And **is today about one careful image or about volume?** That decides how much model you want to pay for. Answer those two and the tables above fill in the rest. [▸ Try these models on PixAI](https://pixai.art/generator/image?utm_source=blog) --- _Source: [https://blog.pixai.art/en/pixai-models-guide/](https://blog.pixai.art/en/pixai-models-guide/) · Part of [PixAI Blog](https://blog.pixai.art/) · Index: [/llms.txt](https://blog.pixai.art/llms.txt)_